Healthcare AI Daily Briefing

Research & Evidence

ML adaptive test reproduces PHQ‑8 and GAD‑7 scores with 2 items median.

Journal article (Journal of affective disorders)

In a psychiatric outpatient cohort, a machine‑learning CAT using the M5P algorithm delivered a median of two items per instrument — cutting 66‑71% of questions — while matching full‑scale scores with r = 0.996‑0.997, R² = 0.986‑0.991 and MAE < 0.62 in 10‑fold cross‑validation. Classification agreement at 6‑month and 1‑year follow‑up was κ = 0.85‑0.93 for PHQ‑8 and κ = 0.91‑0.96 for GAD‑7, and the adaptive versions reliably detected symptom reductions at both time points. [1]

Industry & Products

Pathos raises $237 M Series E at $3 B valuation, led by Bristol Myers Squibb.

AI drug developer Pathos announced a $237 million Series E round led by Bristol Myers Squibb, valuing the company at about $3 billion. The firm is expanding from oncology into neurology, starting with Alzheimer’s through a BMS partnership and new Tempus AI datasets. [2]

Operations & Workflow

Interview: inpatient AI agent fielding observation‑versus‑inpatient status questions.

Podcast from Claims Denied: A Hospitalogy Podcast

The interview reports clinicians using the inpatient AI agent to ask whether a patient meets inpatient versus observation criteria, a decision that drives DRG assignment, fee‑for‑service CPTs via E/M coding, and risk adjustment. [3]

Operations & Workflow

Pairing decision-only AI with an LLM reduces framing‑driven decision changes by about 81% in development and 42–48% on held‑out cases.

Preprint, not peer reviewed

Across 2,134 clinical case presentations and 2,483,222 model calls, adding labels or pressure cues shifted 13.4 more decisions per 100 than neutral wording for the decision-only models Jev and OpenAI Decisions and 14.0 for GPT‑6 Luna. Pairing each decision model with the LLM cut that susceptibility to 2.6 per 100 — roughly 81% lower in development and 42% to 48% lower on held‑out reports — while triage accuracy stayed near 81% to 82% and the combined workflow cost at most US$0.77 per 1,000 entries. [4]